Why would you combine dense and sparse vector search instead of relying on dense embeddings alone?
Hybrid retrieval combines semantic generalization with exact lexical matching
I would combine them because dense embeddings and sparse retrieval fail in different ways. Dense embeddings are excellent at semantic similarity and paraphrases, but they can weaken exact matching for names, SKUs, error codes, abbreviations, model numbers, or rare domain terminology. Sparse retrieval preserves strong lexical evidence for those cases. Running both and fusing their candidate lists gives the system a broader recall envelope: dense handles meaning while sparse handles exact or highly discriminative terms. The trade-off is extra indexing, storage, and query work. A common mistake is assuming the dense model has learned every identifier exactly; in production search I want empirical evidence before removing lexical retrieval.
- 1
Dense retrieval is usually strongest for semantic paraphrases such as "laptop suitable for travel" and "portable notebook computer".
- 2
Sparse retrieval is often strongest for exact tokens such as "RTX 5090", "ERR_CONNECTION_RESET", or a product SKU.
- 3
Trade-off: hybrid search usually improves recall and robustness but increases index/storage footprint and query complexity.
- 4
Fusion method matters because dense and sparse scores can have different scales; rank-based fusion such as RRF avoids relying on directly comparable raw score magnitudes.
Scenario Questions
0-2 years experience
- 1A user searches for an exact product code and dense retrieval returns a similar but incorrect product. Which retrieval signal is missing?
- 2A user writes a vague natural-language query with synonyms. Why might dense retrieval help more than keyword matching?
2-5 years experience
- 1Your search quality is good for prose queries but poor for SKUs and model numbers. How would you change the retrieval architecture?
- 2Why would you avoid assuming that dense and sparse scores can simply be added without normalization or calibration?
5-8 years experience
- 1Design a hybrid retrieval path for an e-commerce catalog with product titles, descriptions, and exact manufacturer part numbers.
- 2You introduce sparse retrieval and recall improves but latency increases by 40%. How would you determine whether the quality gain justifies it?
8+ years experience
- 1How would you design an evaluation framework that separates semantic, lexical, and hybrid retrieval gains across multiple query classes?
- 2A business wants one global ranking formula for dense and sparse results. Would you choose score normalization, RRF, or learned fusion, and why?
Follow-up Questions
- Why is rank fusion often safer than directly adding dense and sparse scores?
- Which query types are most likely to benefit from hybrid retrieval?